Slashdot Mirror
Tesla Autopilot Crisis Deepens With Loss of Third Autopilot Boss In 18 Months (arstechnica.com)
An anonymous reader quotes a report from Ars Technica: It is no secret that Tesla's Autopilot project is struggling. Last summer, we covered a report that Tesla was bleeding talent from its Autopilot division. Tesla Autopilot head Sterling Anderson quit Tesla at the end of 2016. His replacement was Chris Lattner, who had previously created the Swift programming language at Apple. But Lattner only lasted six months before departing last June. Now Lattner's replacement, Jim Keller, is leaving Tesla as well.
Keller was a well-known chip designer at AMD before he was recruited to lead Tesla's hardware engineering efforts for Autopilot in 2016. Keller has been working to develop custom silicon for Autopilot, potentially replacing the Nvidia chips being used in today's Tesla vehicles. When Lattner left Tesla last June, Keller was given broader authority over the Autopilot program as a whole. Keller's departure comes just weeks after the death of Walter Huang, a driver whose Model X vehicle slammed into a concrete lane divider in Mountain View, California. Tesla has said Autopilot was engaged at the time of the crash. Tesla has since gotten into public feuds with both Huang's family and the National Transportation Safety Board, the federal agency investigating the crash. "Today is Jim Keller's last day at Tesla, where he has overseen low-voltage hardware, Autopilot software and infotainment," Tesla said in a statement to Electrek. "Prior to joining Tesla, Jim's core passion was microprocessor engineering, and he's now joining a company where he'll be able to once again focus on this exclusively."
Keller was a well-known chip designer at AMD before he was recruited to lead Tesla's hardware engineering efforts for Autopilot in 2016. Keller has been working to develop custom silicon for Autopilot, potentially replacing the Nvidia chips being used in today's Tesla vehicles. When Lattner left Tesla last June, Keller was given broader authority over the Autopilot program as a whole. Keller's departure comes just weeks after the death of Walter Huang, a driver whose Model X vehicle slammed into a concrete lane divider in Mountain View, California. Tesla has said Autopilot was engaged at the time of the crash. Tesla has since gotten into public feuds with both Huang's family and the National Transportation Safety Board, the federal agency investigating the crash. "Today is Jim Keller's last day at Tesla, where he has overseen low-voltage hardware, Autopilot software and infotainment," Tesla said in a statement to Electrek. "Prior to joining Tesla, Jim's core passion was microprocessor engineering, and he's now joining a company where he'll be able to once again focus on this exclusively."
Pretty soon Elon will be wearing all the hats. Or maybe he'll assign his cyborg dragon to fill some of the roles.
LIDAR is the way to go and both Google and Apple know this.
The problem is, puck-sized LIDAR systems, as seen in 8-packs on the Apple dev car, cost 8000 a piece and that is why Testa uses cheapo-cams and parking radar.
I suspect Tesla's method of using less hardware will be the main path in 15 years for autonomy, once we have car to car communications and car to traffic control communications as standard equipment in every vehicle and bugs worked out.
All the kinks worked out and equipment in every vehicle on the road in the next 15 years. Elon, is that you?
Some cars with humans or lasers can communicate safe passages and routes through construction, and other road conditions (wet/ice/snow...) And lesser equipped cars can then navigate recently validated routes more safely.
I'm sure you're not saying that a ~4,000 pound hunk of metal being propelled at ~60mph can make safe operational decisions based not on current environmental conditions but on environmental conditions that existed at some point before it reached a given area.
I suspect Tesla's method of using less hardware will be the main path in 15 years for autonomy, once we have car to car communications and car to traffic control communications as standard equipment in every vehicle and bugs worked out. Some cars with humans or lasers can communicate safe passages and routes through construction, and other road conditions (wet/ice/snow...) And lesser equipped cars can then navigate recently validated routes more safely.
But now the NTSB is very likely going to step into national standards for autonomy, and it doesn't appear Tesla is ready to meet the likely minimum standard, such as redundant navigation (operate without GPS, or without optical recognition.) and redundant systems.
Hardware only gets cheaper, the future of self-driving cars is more and redundant sensors. And no car is going to rely on another car to tell it what is a safe route.
I stole this Sig
Contrary to some bad reporting of late, the Model 3's AP systems are redundant - both onboard redundancy, and redundancy in the mechanical assist systems. And if AP loses a sensor or a sensor is obstructed, AP is disabled until it can be fixed.
No matter how kind you are, German children are kinder.
It just like airplane autopilot, except airplane autopilot development bosses didn't all quit.
Also, the NTSB does not set standard for autonomy. The NTSB is an investigation board. The NHTSA sets standards in the US.
No matter how kind you are, German children are kinder.
Humans seem to get by ok with minimal hardware. Two cameras and one massively parallel, but extremely buggy and unreliable CPU that completely fabricates a lot of what you see based on its best guess.
I'm surprised that Tesla doesn't already have car-to-car comms for their own vehicles
Pain is merely failure leaving the body
Stop calling it autopilot, rename it to "drive assist".
#DeleteFacebook
Radar would be a problem if you were using 10kW pulsed kind of powers but for different radar technology (FMCW) this can be overcome. FMCW is unsuitable because of having multiple units operating in one area but there is pulse compression radar which is code-able and for the kind of range you need (probably 250m) can be low power. The way these radars work is to put the power on the target over time rather than instantaneous or pulsed, however pulse compression is still a pulse it is spread over a much greater time period. Radar is better than LIDAR at some things, particularly fog or difficult optical environments.
You can't handle the truth! - Because I don't post left all my comments get modded down, bye bye Karma.
Neither Keller nor Lattner is a huge loss.
Lattner wasn't there long and he was out of his depth. Keller is a talented hardware designer but what Tesla and all self-driving companies need is software prowess. Losing Sterling Anderson surely hurt, losing Andrej Karpathy would be a big frickin' deal.
This is not that.
Pain is merely failure leaving the body
But, all the same, I wish they never got into this auto pilot self driving thing. It is a distraction from getting the affordable electric car done. That is the most important thing to get done.
sed -e 's/Chuck Norris/Rajnikant/g' joke > fact
... with several billion years worth of AI training data
In 15 years?
The average car age on the road USA is approaching 12 years. in 2003 is was 9.7 years, in 2016 it was 11.6 years.
In 15 years the average car will be similar to what's available now. Most cars won't have self-driving capability or car to car communication.
There are plenty of cars from the 70's, 80's and 90's on the roads today. In 15 years there still will be. Maybe fewer from the 80's as they start turning to rust.
Maybe in 150 years.
Sure. Then we try and abstract all of that into some electronics and introduce a whole new set of errors and biases. Some predictable, some unforeseen.
Humans are allowed to drive, and generally do so pretty well, even if one "camera" is defective or inoperative.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey
FYI, the Vehicle-to-Vehicle (V2V) system is in some cars today, by 2020 all new cars must have them.
Bigger to me is these communicate to traffic lights, so auto systems in vehicles will get alerts for things like light changes, crashes, weather and temperature. So it is likely a intersection that commonly has problems with Ice will alert.
if AI drivers perform the same, or even half as bad as humans, they would be forbidden.
what I am saying is : humans are bad drivers
And what everyone else is saying is: Tesla's "auto pilot", Uber's full-fledged self driving cars, etc. are worse. Much worse. They're parlor tricks that work in a rigged scenario at best. They're decades away from being as safe and flexible as even a drunk human.
And no car is going to rely on another car to tell it what is a safe route.
That's actually exactly what Autopilot does. In traffic it follows the car in front. The lead car turns blue on the display when the Tesla is locked on to it.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
I can't wait to read about 'lemming mode' megacrashes!
You can pick and adjust your follow distance. Some people choose a short follow distance. I personally wouldn't.
No matter how kind you are, German children are kinder.
Aha. Like there is no other living or driving thing around.
No animals or humans on bikes traversing your lane.
No bridges falling between both cars.
No one throwing stones from non collapsing bridges.
No tire exploding, ever.
No water, ice, fog in your path, ever.
No confusing road markings
No sun blinding cameras
No rain
Just pretend all those things do not exist.
aaaaaaa
The most frightening part is that Tesla's Autopilot, in addition to relying on less hardware compared to more autonomous vehicle like Google's (Level 2 vs. Level 3/4), is even relying on less hardware than LESS autonomous cars.
(Tesla's Level2 is relying on a single Camera,
Whereas Level1 like Volvo add laser lidar in addition to camera,
and several brands including mercedes have been adding stereo cameras for quite some time).
But on the scale of the price of the whole car (I mean, on scale on the price of the whole battery to which the strap a free car as a bonus~ :-P ) does reducing the hardware has really a significant impact ?
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
You're talking about V2V? Yes, they have spectrum and some implementations. They don't even seem to have a single agreed to or dictated protocol that would allow, for example, BMWs to talk to Cadillacs. And no, there is no current Implement_by-date although several have been proposed. NHSTA proposed a standard at the end of 2016, but it seems to have quietly died in late 2017
Not that I'm opposed to V2V? Seems like a perfectly OK idea actually if everyone can be persuaded to speak the same digital language.
But what they have does not seem remotely what will actually be needed to support autonomous vehicles.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey
Maybe that's just the problem here? If I've guy is the boss all the time he will be made responsible for the people who die due to any bug in the autopilot, because they should have seen it. If the boss keeps changing they can always blame someone else.
Maybe drivers who don't pay attention to obatacles as they head directly towards them, shouldnt drive.
I suspect Tesla's method of using less hardware will be the main path in 15 years for autonomy, once we have car to car communications and car to traffic control communications as standard equipment in every vehicle and bugs worked out.
Will _all_ the bugs be worked out before, or after Pagani deliver my flying unicorn.
Calling someone a "hater" only means you can not rationally rebut their argument.
I here the sound of taxes going up for the sake of autonomous driving.... again.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
> Most cars won't have self-driving capability or car to car communication.
Doesn't take most, just takes enough. Vehicle to Vehicle (V2V) was introduced to new cars 2 years ago, and will likely be in all news cars in 2 years. Driver assist (lane assist, emergency braking...) with NHTSA is already recommending CIB, DBS and PAEB braking, and was on a path to make them mandatory, but the current anti regulation path in the US has likely put a delay to that.
With these style of cars, we are a few system updates away from converting them to be part of a smart grid. So 15 years may be optimistic with a deregulation movement likely to slow things for the next 3-7 years. But if we had a hard regulatory push away from petrol cars in 10 years, we could be 90% smart cars in 15-20 years.
It's so minimal that it takes massive amounts of hardware to even poorly replicate specific aspects of its function.
> make safe operational decisions based not on current environmental conditions but on environmental conditions that existed at some point before it reached a given area.
You probably don't realize we all do that. Not exclusively from the past, but it is important. When I drive to work in freezing temperatures, I have a past knowledge making it safer. I know if it rained/snowed/sleeted, I smell see the affects of water to know where it may flood or freeze, I see it on the wipers, I know what roads are high traffic and were salted, and had high traffic to either maintain a clear surface (or were more likely to turn snow into ice, and to avoid.) I know to look out for road sections that are shaded from sun, or bridges.
Similar with looking at tracks of other cars, and have to trust the cars around me, looking way ahead for brake lights over hills...
If a car system assumes the worst, and thus doesn't get help from off car data, they will be driving 20 mph on all paved roads in freezing temperatures. If it drives normally in freezing temps, and doesn't know these special details, it will not be able to stop on the ice in time for intersections, or go into ice in a shaded corner sailing off the road.
If it blindly follows are car in front, it may not know that car is a dump truck into construction, or has studded tires on ice, or more clearance for deep snow, or if a stock Tesla follows to close to something like a corvette that stops from 60 in under 100 feet (stock model Tesla is around 180 feet) not having the same capabilities are things we can know, but the tesla wont without communications.
The vehicles have cameras, radar, lidar and they integrate this information into a single stream. The have a "bullshit detector" known as a Kalman filter to only accept accurate information. Any data that conflicts is discarded. That information comes out as a cloud of points and has to be processed as road surface, road signs, pedestrians, vehicles, obstacles. Those are the easy ones. Then there is ice, snow, water puddles, reflections, vehicles with razzle-dazzle adverts (buses with a collage of road signs), street art (some cities have murals painted on concrete walls. Classic example is the hole-knocked-in-the-wall with a paradise of green meadows and blue sky on the other side). On the rare event side, there are sinkholes, overturning trucks.
One way is just to have the vehicle trained to match the current camera view with what should be done using a training set of video frames.
Vintage computer adverts: http://www.vintageadbrowser.com/computers-and-software-ads
You probably don't realize we all do that. Not exclusively from the past, but it is important. When I drive to work in freezing temperatures, I have a past knowledge making it safer. I know if it rained/snowed/sleeted, I smell see the affects of water to know where it may flood or freeze, I see it on the wipers, I know what roads are high traffic and were salted, and had high traffic to either maintain a clear surface (or were more likely to turn snow into ice, and to avoid.) I know to look out for road sections that are shaded from sun, or bridges.
Similar with looking at tracks of other cars, and have to trust the cars around me, looking way ahead for brake lights over hills...
I get all that, but every single one of those examples involves your immediate senses in combination with your general knowledge of your surroundings, and you make your operational decisions based on the least common denominator of the two. In no circumstance are you squeezing your eyes shut to what you're actually seeing and feeling and instead guessing at your environmental conditions based on some weather/traffic report you heard on the radio 5 minutes ago. And that's ultimately what OP was proposing -- reliance on stale environmental data in an effort to reduce sensor costs.
If a car system assumes the worst, and thus doesn't get help from off car data, they will be driving 20 mph on all paved roads in freezing temperatures.
Good. Hazardous driving conditions dramatically amplify the cost of mistakes. If that's all the faster they can safely go based on what they're able to sense of the real world around them, it'll be even clearer they're not ready for prime time.
> And no car is going to rely on another car to tell it what is a safe route.
Not exclusively, but often it will either by reliant on other cars, or your going to be by far the slowest car to the point of causing a crash.
Almost every day, usually multiple times a day, we go over hills and around corners where we cannot see far enough to see something in time to stop before hitting it. We are either trusting luck, or looking at the cars in front/around, did they hit their brakes? Especially in freezing temperatures after weather, in the morning. We often trust enough cars have driven to keep clear ice, or are driving the usual speeds. Or if the roads are bad, enough cars back up to slow us. Or that we know where the usual risks for issues are.
Some hardware gets cheaper some of the time. I've noticed for instance that the old-fashioned mechanical watches that were the staple of the last century have not gotten any cheaper at all. Recently the price of high-end GPUs has skyrocketed.
Mass production of some goods can cause them to become cheaper but this is by no means a foregone conclusion.
How many billion miles travelled per day? The relevant statistic is the number of fatalities per billion miles (or km) traveled. See here. We don't know for sure the relevant statistics for driver-assisted vehicles.
There is a very good chance that driver assistance is going to improve the fatalities statistics, but this needs to be done correctly.
> And that's ultimately what OP was proposing -- reliance on stale environmental data in an effort to reduce sensor costs.
No, I am OP, and specifically stated Tesla will need V2V communications. Only with no cameras does it still need to be able to navigate for short periods safely in the event of a system failure in the vision system while at speed it cannot have the camera/radar lose power, or get hit buy a mud shower, then stop in traffic, unable to navigate out of the way to a safe location. Same with loss of GPS, it cannot rely on just a GPS model, it must have exact positions of visual marker to navigate short distances with no GPS (such as laser data would provide.)
And you will be the cause of the crash (maybe not legally at fault) if yours is the only car that slows to 5 mph for a corner of unknown navigability. A car with Tesla's sensors will not be able to negotiate at speed several situations, that a laser car can, such as snow covered signs or lane markers, or tunnels/canyons/etc where GPS is lost. But the Tesla can follow/compare it's visual radar data to a recent laser car's 3d model, and then go to those places safely.
FYI, the Vehicle-to-Vehicle (V2V) system is in some cars today, by 2020 all new cars must have them.
Where did you get the 2020 date? Their website states that they have made no such mandate yet:
https://www.nhtsa.gov/press-releases/v2v-statement
"The Department of Transportation and NHTSA have not made any final decision on the proposed rulemaking concerning a V2V mandate."
Just sayin': a few years ago, AI was supposedly "decades away" from beating the best Go players.
Speaking of NHTSA, I was at a conference a month or so ago that was jam packed full of futurists who were under the impression that Uber, Google, and Tesla were just a couple years away from launching saturation-ready fully autonomous vehicles. Then David Friedman, former interim head of NHTSA, gets up to speak in the keynote and burst their bubbles. He basically said that AVs have got to be nearly perfect or (1) people will sue the companies for everything their worth when an injury/fatality happens and (2) the tech is nowhere near affordable or ready to meet this requirement. The crowd went silent.
Of course 100% reliability will not be the standard, because that's never acheivable.
> Where did you get the 2020 date?
Thanks for that, It was in a article from around January 2016. posted to /. so likely just a planned date, as yours is newer.
Until it finds a collapsed crash barrier to 'follow'.
John McAfee 'It was like that time I hired that Bangkok prostitute; to do my taxes, while I fucked my accountant'
Sure they have. Chinese Rolexes have never been cheaper.
John McAfee 'It was like that time I hired that Bangkok prostitute; to do my taxes, while I fucked my accountant'
And a light doesn't change just after the first one passes it.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
Losing three seems like carelessness.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
And, fortunately, cars never go off the road into a ditch.
"When you have eliminated the unacceptable, whatever is left, however improbable, must be the truthiness" - Holmes
That should be fine as long as it keeps a safe distance.
Or as long as it doesn't plunge 200 feet into a ravine.
Some hardware gets cheaper some of the time. I've noticed for instance that the old-fashioned mechanical watches that were the staple of the last century have not gotten any cheaper at all.
The cheap old-fashioned mechanical watches of the last century vanished from the market, replaced by cheap digital watches, smartphones, and fitbits.
The mechanical watches available now are more jewellery than watch.
Recently the price of high-end GPUs has skyrocketed.
That's mostly a supply shortage caused by cryptocurrency miners. Though I'm curious what will happen if mining stays profitable longterm.
Mass production of some goods can cause them to become cheaper but this is by no means a foregone conclusion.
There are exceptions, but I see no reason that car sensors would get more expensive when self-driving cars become mainstream.
I stole this Sig
Then we'd better hope Plod doesn't nab us, because we're committing an offence.
https://www.bcc-wv.com/blog/20...
https://en.wikipedia.org/wiki/...
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
I think your numbers are slightly off, unless the first drivers were bacteria.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
It may be a surprise for some, but there is a theory we all evolved from single celled organisms.
The human brain wasn't created by an almighty being.